Use case
Finance operations agents
Finance teams spend hours reading invoices, matching POs and chasing exceptions. A Dezifi finance agent reviews every document, flags only what needs a human, and routes everything else through.
What you'll learn
- How to extract structured data from PDFs and emails
- Which checks to encode as policies vs. agent reasoning
- How to build a reviewer queue with full audit trail
- How to keep the agent honest with deterministic rules
The agent design
Finance is high-stakes. The agent should narrow the queue, never approve money movement. Determinism beats cleverness here.
- 1
LLM choice
Claude Sonnet or GPT-4o with vision for invoice OCR. Pair with deterministic regex / schema validators to keep extraction grounded. - 2
Tools
NetSuite, QuickBooks or Xero (read POs, post journal entries with approval), SAP or Oracle if applicable, Gmail (read AP inbox), DocuSign, Slack for reviewer pings. - 3
Guardrails
PII and PAN masking in logs. Any journal entry, payment release or vendor master change routes to approval. Hard cap on dollar amount the agent can process unattended (e.g., zero). - 4
Workflow shape
Trigger: invoice email lands in AP inbox. Step 1: extract structured fields. Step 2: match to PO and receipt. Step 3: run policy checks (duplicate, vendor allowlist, amount limit). Step 4: route exceptions to reviewer queue; clean invoices to auto-post (still gated).
Tools to connect
- NetSuite / QuickBooks / Xero — ERP for POs, invoices, journal entries.
- Gmail or Outlook — AP inbox monitoring.
- DocuSign — pull executed contracts for vendor validation.
- Slack — reviewer pings, escalation channel.
How to set this up in Dezifi
- 1
Connect the ERP
Integrations → NetSuite (or your ERP) → service-account OAuth. Scope to the AP-only sub-account. - 2
Connect the AP inbox
Integrations → Gmail → OAuth on the shared AP mailbox. Use a label-watch on incoming invoices. - 3
Create the agent
New Agent → "Invoice Triage Agent". Attach ERP, Gmail, Slack. Use Claude Sonnet with vision enabled. - 4
Encode rules as policies
Add policies: vendor must exist in vendor master; invoice total must match PO ±2%; duplicate invoice number within 90 days triggers exception. These run deterministically — not via LLM judgment. - 5
Build the reviewer queue
Workflow: agent processes invoice → policy engine evaluates → exceptions route to Approvals queue with full context (extracted fields, matched PO, policy that failed). Reviewer approves or rejects; everything is logged.
Frequently asked questions
- Can the agent post journal entries by itself?
- Configure it to. We recommend not. The platform supports a hard cap that requires human approval for any movement of money. Start there and only loosen when you have weeks of clean traces.
- How accurate is invoice extraction?
- For typed PDFs, modern vision-LLMs are 95%+ accurate on standard fields. Handwritten or low-resolution scans drop accuracy — those should always route to the reviewer queue.
- Does this work for non-USD currencies and tax regimes?
- Yes. Currency, tax codes, and locale are part of the structured extraction. Encode regime-specific rules (GST, VAT) as policies.
- How do we satisfy auditors?
- Every run is fully traced. Monitor and Analytics export complete audit logs — LLM input/output, tool calls, reviewer decisions — for every invoice.